Xiaojian Liao

Xiongwei Zhu, Xiaojian Liao, Tianyang Jiang et al.

Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under llama.cpp on Jetson Orin NX, corresponding to a 1.77-1.99$\times$ decode speedup across diverse workloads. Checkpoints and usage instructions are available at https://github.com/BUAA-OSCAR/ReMoE.

Jiashun Suo, Xiaojian Liao, Limin Xiao et al.

Large language models like GPT-4 are resource-intensive, but recent advancements suggest that smaller, specialized experts can outperform the monolithic models on specific tasks. The Collaboration-of-Experts (CoE) approach integrates multiple expert models, improving the accuracy of generated results and offering great potential for precision-critical applications, such as automatic circuit board quality inspection. However, deploying CoE serving systems presents challenges to memory capacity due to the large number of experts required, which can lead to significant performance overhead from frequent expert switching across different memory and storage tiers. We propose CoServe, an efficient CoE model serving system on heterogeneous CPU and GPU with limited memory. CoServe reduces unnecessary expert switching by leveraging expert dependency, a key property of CoE inference. CoServe introduces a dependency-aware request scheduler and dependency-aware expert management for efficient inference. It also introduces an offline profiler to automatically find optimal resource allocation on various processors and devices. In real-world intelligent manufacturing workloads, CoServe achieves 4.5$\times$ to 12$\times$ higher throughput compared to state-of-the-art systems.

Shudong Wang, Xinfei Wang, Chenhao Zhang et al.

Multi-task reinforcement learning (MTRL) seeks to learn a unified policy for diverse tasks, but often suffers from gradient conflicts across tasks. Existing masking-based methods attempt to mitigate such conflicts by assigning task-specific parameter masks. However, our empirical study shows that coarse-grained binary masks have the problem of over-suppressing key conflicting parameters, hindering knowledge sharing across tasks. Moreover, different tasks exhibit varying conflict levels, yet existing methods use a one-size-fits-all fixed sparsity strategy to keep training stability and performance, which proves inadequate. These limitations hinder the model's generalization and learning efficiency. To address these issues, we propose SoCo-DT, a Soft Conflict-resolution method based by parameter importance. By leveraging Fisher information, mask values are dynamically adjusted to retain important parameters while suppressing conflicting ones. In addition, we introduce a dynamic sparsity adjustment strategy based on the Interquartile Range (IQR), which constructs task-specific thresholding schemes using the distribution of conflict and harmony scores during training. To enable adaptive sparsity evolution throughout training, we further incorporate an asymmetric cosine annealing schedule to continuously update the threshold. Experimental results on the Meta-World benchmark show that SoCo-DT outperforms the state-of-the-art method by 7.6% on MT50 and by 10.5% on the suboptimal dataset, demonstrating its effectiveness in mitigating gradient conflicts and improving overall multi-task performance.

Tianhao Cai, Liang Wang, Limin Xiao et al.

With the rapid development of DNN applications, multi-tenant execution, where multiple DNNs are co-located on a single SoC, is becoming a prevailing trend. Although many methods are proposed in prior works to improve multi-tenant performance, the impact of shared cache is not well studied. This paper proposes CaMDN, an architecture-scheduling co-design to enhance cache efficiency for multi-tenant DNNs on integrated NPUs. Specifically, a lightweight architecture is proposed to support model-exclusive, NPU-controlled regions inside shared cache to eliminate unexpected cache contention. Moreover, a cache scheduling method is proposed to improve shared cache utilization. In particular, it includes a cache-aware mapping method for adaptability to the varying available cache capacity and a dynamic allocation algorithm to adjust the usage among co-located DNNs at runtime. Compared to prior works, CaMDN reduces the memory access by 33.4% on average and achieves a model speedup of up to 2.56$\times$ (1.88$\times$ on average).